IntElicit: Eliciting and Assessing Contextualized Creativity via Dialogue Policy Optimization

Researchers introduce IntElicit, an AI framework that uses adaptive dialogue policy optimization to assess creativity in interactive environments while filtering out confounding factors like domain knowledge gaps. The approach shows promise in revealing creative potential that traditional static assessments miss, particularly relevant for AI-mediated learning contexts.

IntElicit addresses a fundamental challenge in creativity assessment: distinguishing genuine creative ability from performance confounded by cognitive proficiency or lack of engagement. The framework operates as an adaptive AI interviewer that scaffolds interactions across multiple turns, eliciting creative responses while maintaining participant agency. This represents a meaningful shift from static assessment methods toward dynamic, contextualized evaluation aligned with how creative work actually occurs in human-AI collaborative environments.

The research tackles sparse reward problems endemic to open-ended dialogue systems through decomposed process rewards that incentivize pedagogical prompting rather than answer generation. This mechanic prevents reward hacking—where systems optimize for superficial goal achievement rather than genuine learning facilitation. The distinction matters substantially for educational AI applications, where the goal involves drawing out participant reasoning rather than demonstrating optimal solutions.

The experimental validation through both simulated participants and a 64-person human study provides evidence that interactive elicitation reveals capabilities static approaches overlook. For the AI and educational technology sectors, this work signals growing sophistication in designing dialogue systems that maintain high ecological validity while reducing assessment bias. The implications extend beyond academia into enterprise learning platforms and human-AI collaboration tools, where accurate capability assessment directly impacts decision-making and talent development.

Future applications likely involve integrating similar elicitation mechanisms into productivity tools, creative collaboration platforms, and professional development systems. The framework's emphasis on process-level rewards offers transferable patterns for other dialogue-based assessment contexts beyond creativity evaluation.

• →IntElicit uses adaptive dialogue policy optimization to separate creative ability from domain knowledge and engagement confounders.
• →Decomposed process rewards prevent systems from dictating answers and instead reward prompts that draw out participant reasoning.
• →Interactive elicitation revealed creative potential that static assessment methods systematically missed in controlled studies.
• →The framework aligns with contemporary creative practice in tool-mediated and human-AI interactive environments.
• →64-participant human subject study demonstrated IntElicit outperforms expert-designed baseline assessment approaches.